Station Distance Dir Tier PM2.5 Predicted Level Lead Time
📡
Select a city and click Fetch Live Data or Run Demo
Click fetch or demo to see station data on the map
Alert Levels: Low Moderate High Very High Extreme Target City No Data
C.L.E.A.R. System

Before the Sky Turns Orange

The C.L.E.A.R. System: Canadian Lead-time Early Air Response

Hugo Bui & Ryan Zander — University of Toronto Schools

Providing 6–48 Hours of Advance Warning for Wildfire Smoke

Project Summary

C.L.E.A.R. is a wildfire PM2.5 early warning system for Toronto, Edmonton, Montréal, and Vancouver that repurposes existing NAPS air quality stations located 100–600+ km away to provide 6–48+ hours of advance warning before dangerous smoke arrives.

36M+
Hourly Observations
218
Stations Identified
4 Cities
Covered
20 yrs
Study Period (2003–2023)
97.9%
Detection Rate
0%
False Alarm Rate

Research Question

"Can hourly PM2.5 readings at distant monitoring stations predict a city's PM2.5 levels with enough lead time to issue meaningful public health warnings?"

Purpose

Develop an early warning system that uses existing NAPS and U.S. EPA monitoring infrastructure — stations located 100–600+ km upstream — to detect approaching wildfire smoke plumes and issue colour-coded health alerts 6–48 hours in advance.

The system targets four major Canadian cities: Toronto, Edmonton, Montréal, and Vancouver, covering diverse geographic and smoke exposure patterns.

Data Sources

NAPS Network (Canada)

National Air Pollution Surveillance Program

Hourly PM2.5 measurements. 84% of stations concentrated in BC, AB, ON, and QC. Dense regional coverage enables upstream smoke detection.

U.S. EPA AQS / AirNow

Air Quality System & AirNow

Daily PM2.5 data from U.S. border stations (NY, PA, VT, WA, OR). Critical for cities near the U.S. border.

Analysis Method

  • Study period: 2003–2023, wildfire season (May–September)
  • Correlation analysis between each remote station and the target city's PM2.5
  • Regression model: PM2.5city = slope × PM2.5station + intercept
  • Station selection criteria: R ≥ 0.30, P < 0.001, N ≥ 100 observations
  • Tier classification: Tier 1 (>250 km, 12–48 hr lead) and Tier 2 (100–250 km, 6–18 hr lead)
  • AI validation: Claude, ChatGPT, Gemini, and Julius AI all produced mathematically equivalent regression results; hand-verified

Historical Validation — Confusion Matrix

The system was validated against all historical smoke events (2003–2023):

City True Pos. False Neg. True Neg. False Pos. Detection False Alarm Status
Vancouver140100 100%0%Perfect
Edmonton9150 90%0%1 Failure
Toronto100100 100%0%Perfect
Montréal140100 100%0%Perfect
Total471350 97.9%0%83 events

Edmonton's single missed event was due to a NNW monitoring gap — smoke arrived from an unmonitored direction.

Real-World Example — Toronto, June 2023

During the catastrophic Quebec wildfire event:

DateStationDistancePM2.5ThresholdAlert
June 3Parry Sound187 km NNW 196 µg/m³>139.3 EXTREME
June 6Cornwall387 km ENE 160 µg/m³>99.3 VERY HIGH
June 6Toronto 97 µg/m³(Actual) HIGH

Burn Area Analysis

Analysis of the Canadian National Fire Database shows statistically significant increases in national burn area over the study period. The 2023 season burned over 18 million hectares, more than double any previous year.

18M+ ha
2023 Record Burn
2003–2023
Analysis Period
Significant
Upward Trend

AI Platform Validation

Regression analyses were independently replicated across four AI platforms to ensure mathematical accuracy:

Claude
Anthropic
ChatGPT
OpenAI
Gemini
Google
Julius AI
Data Analysis

All four platforms produced mathematically equivalent results, which were then hand-verified against raw data.

C.L.E.A.R. Alert Levels

The system uses a five-tier alert scale based on predicted PM2.5 concentrations:

Alert LevelPM2.5 LevelPublic Health Action Plan
LOW < 20 µg/m³ No precautions needed.
MODERATE 21–60 µg/m³ Sensitive groups (children/elderly) avoid strenuous activities.
HIGH 61–80 µg/m³ Everyone should reduce physical exertion. N95 or KN95 mask. Keep doors and windows closed. HVAC to recirculate. Run HEPA filter.
VERY HIGH 81–120 µg/m³ Avoid all outdoor activity. Keep hydrated.
EXTREME > 120 µg/m³ Halt indoor pollution. No frying or sauteing. No vacuuming. No candles. No wood-burning stoves.

How It Works

For each included station, the system uses a simple linear regression:

PM2.5city = slope × PM2.5station + intercept

Alert thresholds at each station are computed by inverting the formula:

station threshold = (alert trigger − intercept) ÷ slope

When a station's live PM2.5 reading exceeds its computed threshold, the corresponding alert is triggered — providing advance warning based on the station's distance and tier classification.

Conclusion

Novel Contribution

First system to repurpose existing NAPS infrastructure specifically for wildfire smoke early warning across multiple Canadian cities.

Proven Reliability

97.9% detection rate (47/48 events) with 0% false alarms (35/35 non-events) across all four cities.

Actionable Results

Colour-coded alerts with specific health recommendations give the public clear guidance hours before smoke arrives.

Scalable Solution

The methodology can be extended to any city with nearby upstream monitoring stations.

Future Work

  • Satellite integration for real-time smoke plume tracking
  • More cities added to the network
  • Live testing during upcoming wildfire seasons
  • Expanded NAPS coverage to address monitoring gaps (e.g., Edmonton NNW)
  • Wind direction/speed incorporation for improved predictions

References

  • Anthropic (2026). Claude AI Platform.
  • Cruz, M. G. et al. (2019). Fire dynamics and behaviour.
  • Environment and Climate Change Canada. National Air Pollution Surveillance (NAPS) Program.
  • U.S. Environmental Protection Agency. Air Quality System (AQS) & AirNow.
  • IQAir / World Health Organization. PM2.5 health guidelines.
  • National Forestry Database. Canadian wildfire statistics.

Send Us Feedback

Have suggestions, found a bug, or want to share how you're using C.L.E.A.R.? We'd love to hear from you.

ryanzander99@gmail.com